AI for customer service: what to automate, what to hand to a person

Customer service in a service business is mostly the same dozen questions asked a hundred ways. When are you coming? What do I need to do before you arrive? Do you service my town? Can I move my appointment? What do I owe? AI handles those questions well when it's given accurate answers. It handles the other conversations badly: the customer whose fence was damaged, the one disputing a bill, the one describing a gas smell. A good setup sorts the two quickly.

Which customer conversations can AI handle well?

Look at your last month of calls, texts and emails and you'll find most of them fall into a few groups:

• Information. Hours, service area, services offered, what to expect at the first visit, how to prepare.
• Intake. Name, address, the problem, photos, preferred times. AI is good at asking the same questions in the same order every time.
• Status. "Is the crew still coming today?" If the assistant can read your schedule, it can answer this without anyone stopping work.
• Rescheduling within rules. Moving an appointment to another open slot of the same length is routine. Moving a job that needs a specific crew or a permit inspection is not.
• Payment links and receipts. Sending a link to pay an invoice the office has already issued.

For each of those, the assistant is only as good as the information you give it. If your service area changed last spring and nobody updated the assistant, it will confidently tell customers the old one.

Keep a simple list of these answers in one place, written in your own words: hours, towns, services, what to expect, how to prepare, how to pay. That list becomes the assistant's source of truth, and it doubles as training material for the next person you hire in the office. When something changes, update the list first and the assistant second, so the two never disagree. If your website already answers a question well, point the assistant at that wording instead of writing a second version that can drift.

Which conversations should go straight to a person?

The handoff itself matters. A good handoff passes a summary so the customer doesn't have to repeat everything, and if nobody is free it promises a callback time and alerts the right person. Our guide on handling customer complaints covers what that person should do next.

How does AI fit on each channel?

Phone. An AI receptionist answers calls your team can't, takes intake details and books or routes the job. It's the channel where speed matters most, and it's covered in depth in our AI receptionist guide and the questions to ask a provider.

Text. Customers increasingly text the business number. AI can draft replies from the customer's job history for a staff member to approve with one tap, or answer simple status questions automatically. Approval-first is the safer starting point. Our guide to two-way texting covers the setup, and business texting from a local number needs carrier registration first.

Website chat. A chat assistant on your site can answer the questions your service pages already answer, collect a request and offer booking. It should never invent policies. Limit it to answers drawn from your own published information.

Email. The best use is drafting: summarize the customer's message, propose a reply, and let a person edit and send. Long emails often contain a complaint buried in paragraph three, which is exactly what a summary helps you spot.

What rules should the assistant follow?

Write them down before launch. A one-page rule sheet covers most of it:

1. Prices. Which prices it may quote (fixed-price services only, if any) and the exact wording for everything else ("We price this kind of work once we've looked at it in person").
2. Dates. Which appointments it may book or move, and which need a person.
3. Policies. Cancellation, deposits, warranty, rain days, quoted word for word from your written policies.
4. Never say. No diagnosis, no safety advice beyond your scripted emergency wording, no discounts, no admissions about damage, no legal or insurance opinions.
5. Escalation. The trigger words and situations that go to a person, and who that person is at each hour of the day.

Then test it the way a customer would. Request a firm number for work you only price after a visit. Ask it to move a job to a day you're closed. Describe an emergency in plain words. Watch what it does when it doesn't know.

Should customers know they're talking to AI?

Yes. Have the assistant say it's an automated assistant for your company, and answer honestly if someone asks whether they're talking to a person. Customers who find out later that they were misled tend to remember it. There's a legal angle in some places as well: California's bot law prohibits using a bot online to talk with a person in California while intending to deceive them about being a bot in order to drive a sale, and says a person who discloses the bot isn't liable under that section. Rules elsewhere may differ, so ask a lawyer how they apply to you. Everything on this page is general information and isn't legal advice.

Calls you place are a separate question. In its February 2024 declaratory ruling, FCC 24-17, the FCC said the TCPA's limits on calls with an "artificial or prerecorded voice" reach the human-sounding voices today's AI produces, which means a business needs the person's consent before such a call goes out, apart from emergencies and specific exemptions. If you want AI to make reminder or follow-up calls, check consent first. Texts carry their own rules; our TCPA guide covers them.

Who is responsible when the AI gets it wrong?

You are. An assistant that answers in your company's name is speaking for your company. If it promises a refund, quotes a wrong price or gives bad advice, customers will hold you to it, and regulators won't accept "the software said it." The FTC's 2024 sweep of AI cases made the same point: ordinary consumer protection law covers AI, with no special carve-out.

Privacy needs the same care. Decide what the assistant can see. It may need a customer's next appointment time. It rarely needs their full payment history, and it should never read card details aloud or share account information with a caller it hasn't verified. If you use a third-party tool, read its data policy: how long it keeps transcripts, whether they're used to train models, and whether you can export or delete them.

Where should you start?

Start with the channel where customers wait longest. For most contractors that's the phone during working hours and after dark, then texts that sit unanswered while the crew is on a roof. Cover those first with tight rules and easy handoffs. Add website chat and email drafting later.

The assistant also works best when it reads from the same records your office uses: one schedule, one customer list, one set of policies. That's the reason we pair an AI receptionist with a custom CRM for many service businesses. When a policy changes, it changes once, and the phone, texts and website all say the same thing.

Frequently asked questions